|From||"Cavallo, Alexander" <firstname.lastname@example.org>|
|To||"Statalist (email@example.com)" <firstname.lastname@example.org>|
|Subject||st: St: glogit|
|Date||Fri, 2 Apr 2004 11:11:56 -0600|
I am estimating a glogit model explaining the percent female in occupations. I have aggregated my data (both dependent and independent variables) using pweights. Because of this, percent female does not equal # females/total # in occupation, but rather sum(female wgts)/sum(all wgts in occupation). My concern is in getting the weights right for WLS. The manual states that the weight for each cell is proportional to 1/(nj*pj*(1-pj)) where nj is the number of observations for the cell and pj is the predicted probability. If I use the unweighted counts for the analysis, then I get the dependent variable wrong with the "correct" number of obs for variance calculation. If I use the weigthed counts, then I have the right dependent variable but nj is about 500 times too big (the average weight is 500).
Here's my idea on solving this:
Create new count variable:
newfemales=wgtfempct * nobs
glogit newfemales nobs x1 x2 etc
where wgtfempct is the weighted % female
nobs is the unweighted # of obs in occ
Is anyone aware of a literature on this? Any comment would be appreciated.
332 South Michigan Avenue
Chicago, IL 60604
(312) 322-0208 voice
(312) 322-0218 fax